Revisiting Unsupervised Local Descriptor Learning

Constructing accurate training tuples is crucial for unsupervised local descriptor learning, yet challenging due to the absence of patch labels. The state-of-the-art approach constructs with heuristic rules, which struggle precisely depict real-world transformations, in spite enabling fast model convergence. A possible solution alleviate problem clustering-based approach, can capture realistic variations and learn more class decision boundaries, but suffers from slow This paper presents HybridDesc, an that learns powerful models convergence speed by combining rule-based approaches construct tuples. In addition, HybridDesc also contributes two concrete enhancing mechanisms: (1) a Differentiable Hyperparameter Search (DHS) strategy find optimal hyperparameter setting so as provide prior (2) On-Demand Clustering (ODC) method reduce clustering overhead without eroding its advantage. Extensive experimental results show efficiently descriptors surpass existing even rival competitive supervised ones.